Chapter 01
A single conversation
Within one interaction, a model can use its context well. When the interaction ends, the context is gone — and with it what was learned, decided and promised.
Concept visuals
How a long-lived system can keep its history: a model at work, a memory that outlives the conversation, and checked rules between them. Scroll through the chapters.
Chapter 01
Within one interaction, a model can use its context well. When the interaction ends, the context is gone — and with it what was learned, decided and promised.
Chapter 02
Events are written to a memory that outlives the conversation. Each record keeps where it came from and when it was known, so the history can be read back later, in order.
Chapter 03
What a model writes is a proposal. Checked rules decide what becomes a record. Everything else is kept, but marked as the model’s own words — never as a fact or a decision.
Chapter 04
Every record keeps when it became known. So the memory can answer two different questions — what is true now, and what was known at a given moment — without guessing.
Chapter 05
When a conclusion changes, the correction is added as a new record that points back to the earlier one. The earlier view stays in the history, so the path between them is never rewritten.
Chapter 06
Models can be replaced. The history and the checked rules belong to the system, so a new model starts from the same memory — nothing lost, nothing rewritten.
An illustration of principles, not a diagram of our system. Drag to rotate.
Smaller visuals for the ideas behind our research. Tap or click a card to replay it.
Intelligence is not only a sequence of answers. Some problems require coherence across time.
Different problems may require different levels of reasoning effort.
An LLM should be able to revise conclusions when stronger evidence appears.
More computation is not always more intelligence. The value of additional reasoning matters.